WAS/Minimax-h3_Singularity_v2_pruned

MiniMax H3 Singularity ref2va v2, pruned

Pruned and quantized versions of Minimax-h3_Singularity_ref2va_v2_int8.safetensors from WarmBloodAban/Minimax-h3_Singularity, in the same layout as ComfyUI's own pruned MiniMax H3 releases. Load them in ComfyUI with Load Diffusion Model, with weight_dtype left at default.

File	Size	Transformer blocks	Token refiner
Minimax-h3_Singularity_ref2va_v2_pruned_int8_convrot.safetensors	21.0 GB	int8 convrot, byte-identical to the source	bf16
Minimax-h3_Singularity_ref2va_v2_pruned_w6a8.safetensors	16.0 GB	w6a8, group 32	int8 convrot
Minimax-h3_Singularity_ref2va_v2_pruned_w4a8.safetensors	11.8 GB	w4a8, group 16	int8 convrot

What was pruned

The timestep embedder and the full-width adaLN projection in each of the 50 transformer blocks, 13 billion weights in all, are replaced by a 1025 x 8 adaln_t_table and 8-wide adaLN weights: the curve layout of ComfyUI's own minimax_h3_ref2va_pruned_int8_convrot. ComfyUI loads it natively. Every other tensor in the int8 file is byte-identical to the source, and the modulation the pruned layers produce is within 0.025% of the source's at every timestep tested.

The w6a8 and w4a8 files start from the pruned int8 and quantize the transformer block linears and final_layer.video_out, with the token refiner in int8 convrot, as ComfyUI's pruned w6a8 release does.

Comparison

1344 x 768, 124 frames at 24 fps, seed 42, euler, simple, 6 steps, DMAD 4-step LoRA at 0.7, PyTorch attention, minimax_h3_video_vae_int8_convrot and the Qwen3-VL 32B MiniMax H3 text encoder. Mean PSNR of every frame against the source render:

File	PSNR	Render
source v2 int8		source_v2_int8.mp4
pruned int8 convrot	29.1 dB	pruned_int8_convrot.mp4
pruned w6a8	26.2 dB	pruned_w6a8.mp4
pruned w4a8	21.4 dB	pruned_w4a8.mp4

The same source, rendered again with the same settings under a different memory state in ComfyUI, differs from its first render by 24.7 dB, more than the pruned int8 does. Rendered twice under the same memory state, it is identical frame for frame.

source_v2_int8_other_memory_state.mp4

LoRAs without adaLN or timestep-embedder keys apply unchanged, including the source card's recommended minimax_h3_ref2v_turbo_4step_v0.1.

Modified files

These files are modified versions of Minimax-h3_Singularity_ref2va_v2_int8.safetensors (MiniMax H3 Community License Agreement, Section III.2): the timestep path is replaced as described above, and the w6a8 and w4a8 files are re-quantized. Each file's safetensors metadata states the same.

Credits

Fine-tune: WarmBloodAban. Base model: MiniMax H3 by MiniMax.

License

MiniMax H3 Community License Agreement. See LICENSE and NOTICE.